Introduction

The purpose of this analysis is to present the age-gender specific BMI prevalence used in the Dcision Support tool for Child and Adolescence Obesitys model (“the model”).

Nationally representative data from the Australian Bureau of Statistics National Health Survey (NHS) collected in 2007/2008, 2011/2012, 2014/2015, and 2017/2018, Longitudinal Study of Australia Children (LSAC) collected over 8 waves between 2004-2018, and the 2007 National Children Nutrition and Physical Activity Survey are combined and summarized to estimate the prevalence of BMI categories across multiple age groups by gender over time.

The prevalence of each BMI category over time is used to calibrate the model. However, the relatively small number of time points across the three data sources affected the speed and quality of calibrations. To improve calibrations, a predicted line of best fit was created to use in the model objective function. This gives the model a continuous set of values to calibrate off rather than the limited number of time points between 2007 - 2018.

Aims

The aims are to (1) summarise the prevalence of underweight, healthy weight, overweight, and obesity over time across three data sources, and (2) create continuous estimates of BMI prevalence across 2007 - 2018 to be used in model calibration.

Method

The data sources used to summarise BMI prevalence have nationally representative weights used to correct for sampling processes. These weights are used to create weighted BMI prevalence. The prevalence and 95% CI are presented in plots.

When constructing the predicted value of BMI prevalence between 2007 - 2018, it was essential that the bounded nature of summary prevalence (0%-100%) and that the sum of BMI prevalence is 100% was maintained. Multinomial logistic regression was used to ensure that the predicted outputs adhere to these constraints.

It should be noted that using outcome data to both calibrate and validate the SDM is problematic. Ideally, the model should be calibrated on a subset of the data; however, with such little data, the multinomial predictions are sensitive to the removal of the last data point (the data point that could be used for an out-of-sample validation). The use of calibration should be reported as a limitation throughout the reporting of the

Results

Converting age groups

To reduce possible identification, NHS uses Confidential Unit Record File (CURF) data, which reports individuals within 5-year age groups. These reported age groups differ from those used in the system dynamic model and were updated after the 2007 collection. A weighted mean is taken to convert estimates from between age groups; this assumes a uniform distribution of BMI within each age group.

Example The SDM uses a 3-year age cohort for children, e.g. children aged 9-11 years old (9-10, 10-11, 11-12). However, this age group belongs to two NHS data 5-year groups, 5-9 years and 10-14 years. In this situation, 1/3 of the prevalence for 5-9 years olds and 2/3 of the prevalence of 10-14 years are combined to estimate the prevalence of 9-11 year olds.

The table below compares the age groups from the NHS and the SDM; the reported age groups for NHS also differ between 2007 and the remaining collection years.

Table 1 Age conversion between National Health Survey (NHS) and SDM age groups

NHS		SDM
2007	2011-2017
-	Age 0-4	Age 2
		Age 3-5
Age 5-9	Age 5-9
		Age 6-8
		Age 9-11
Age 10-14	Age 10-14
		Age 12-14
Age 15-17	Age 15-17	Age 15-17
Age 18-24	Age 18-19	Age 18-19
	Age 20-24	Age 20-24
Age 25-29	Age 25-29	Age 25-29
Age 30-34	Age 30-34	Age 30-34
Age 35-39	Age 35-39	Age 35-39
Age 40-44	Age 40-44	Age 40-44
Age 45-49	Age 44-49	Age 44-49

Comparing age-group conversions

The following plots compare the age-groups that were converted for each collection year for the National Health Survey (NHS).

2007 In the 2007 NHS, the first age group was the 5-9 year olds; the SDM age group 3-5 year olds (3-4, 4-5, 5-6) was assumed to be the same value as NHS 5-9. SDM 9-11 and 12-14 use weighted means to estimate the prevalence. SDM 18-19 and 20-24 were assumed to be equal to NHS 18-24.

Figure 1: Comparison of NHS and SDM age groups 2007

2011 From 2011 onwards, there was a slight change to the reporting age groups, which more closely aligned with the SDM age groups. The prevalence of SDM 2 years old was assumed to be equal to 0-4 years old (under the uniform assumption). SDM age groups 3-5, 6-8, 9-11, 12-14 were weighted mean prevalence. Age groups 12-14 matched between the two definitions.

Figure 2: Comparison of NHS and SDM age groups 2011

Figure 3: Comparison of NHS and SDM age groups 2014

Figure 4: Comparison of NHS and SDM age groups 2017

Outcome by BMI Category and source

The LSAC and NPAS recorded individuals’ age to the nearest year and were grouped according to the SDM age group definition.

Figures 5 and 6 combine all three data sources (NHS, LSAC, NPAS) and present national trends in BMI prevalence between 2007 and 2018 for each gender.

Figure 5: Males BMI Prevalence over time using multiple sources

Figure 6: Females BMI Prevalence over time using multiple sources

Fitted values

Figures 7 and 8 present the fitted predicted values from the age-gender-specific multinomial regression models. While these values may appear linear, the purpose of using this method is to ensure that fitted values always sum to 1 and will remain bounded between 0 and 1, a feature that can’t be guaranteed using a linear regression model.

Figure 7: Fitted values from multinomial regression modelling, males

Figure 8: Fitted values from multinomial regression modelling, females

Conclusion

This report gives an overview of the methods used to convert age groups for the Australian Bureau of Statistics National Health Survey to align with SDM age groups. The results summarise the prevalence of each BMI category over time between the three main data sources used in the model, ABS NHS, LSAC and 2007 NPAS. To improve performance, predicted values from a multinomial regression model were used to estimate the BMI prevalence values between observed time points. The results of this model create a “line of best fit” that ensures that the bounded nature of BMI prevalence is maintained and that the sum of BMI prevalence equals 100%. The results from multinomial regression models used to produce fitted values used in the model to assist in calibrating the DiSCAO model.

Appendix : Input variables for DiSCAO